pyannote-rs — pyannote.audio as a dependency-isolated RemoteMedia SDK plugin

Hand-rolled AsyncStreamingNode plugin (Path 3 in the RemoteMedia SDK plugin guide) that registers two node types — PyannoteEmbedding and PyannoteSegmenter — by wrapping upstream pyannote-rs 0.3.x.

Why hand-rolled (Path 3)?

Dependency isolation. The dual-emit path (Path 2) would force the plugin's ort version to match the host's workspace pin:

Layer	ort version	Reason
Host (remotemedia-core etc.)	=2.0.0-rc.12	Latest, needed for other nodes
pyannote-rs 0.3.x	^2.0.0-rc.10	Hard-pinned; rc.11 jumped ndarray 0.16→0.17, rc.12 added !Sync operators

Path 2 would resolve these to a single version → broken build. Path 3 sits in its own workspace, depends ONLY on remotemedia-plugin-sdk (NOT remotemedia-core), and pins ort=rc.10 itself. The host links remotemedia-core (and therefore ort=rc.12) and never sees the plugin's ort=rc.10 — they cross the dlopen boundary as opaque abi_stable types.

Use from a manifest

{
  "version": "v1",
  "plugins": ["pyannote-rs@v0.1.0"],
  "nodes": [
    {
      "id": "embed",
      "node_type": "PyannoteEmbedding",
      "params": { "embedding_model_path": "./wespeaker_en_voxceleb_CAM++.onnx" }
    },
    {
      "id": "seg",
      "node_type": "PyannoteSegmenter",
      "params": { "segmentation_model_path": "./segmentation-3.0.onnx" }
    }
  ]
}

The SDK resolver expands pyannote-rs@v0.1.0 to github.com/RemoteMedia-SDK/pyannote-rs, fetches plugin.toml, then falls through to release-manifest.json for the platform-specific prebuilt .so / .dylib / .dll asset.

Status: plugin.toml + source published. Prebuilt release binaries are not yet uploaded — the matrix-build CI workflow is pending. Until then, consumers should either build the cdylib themselves (see below) or use a local-path plugin entry.

Build the cdylib locally

git clone https://github.com/RemoteMedia-SDK/pyannote-rs
cd pyannote-rs
cargo build --release
# → target/release/libpyannote_conflict_plugin.so

Then reference it from your manifest:

{ "plugins": ["./path/to/libpyannote_conflict_plugin.so"] }

What it exports

Node type	Input	Output
PyannoteEmbedding	Audio (16 kHz mono, ~10 s)	JSON { embedding: [f32], dim: int }
PyannoteSegmenter	Audio (16 kHz mono)	JSON { segments: [{start, end}] }

Both nodes require an ONNX model file passed via params (embedding_model_path / segmentation_model_path). See the pyannote-rs README for the model URLs.

License

See LICENSE.md. This plugin reuses RemoteMedia SDK source and is governed by the same RemoteMedia SDK Community License 1.0. The upstream pyannote-rs crate is MIT.